WiDS Datathon 2019

Hello Fast.ai team, We would like to thank you for developing fast.ai over pytorch. It’s really easy to use and very accurate.

More info about WIDS / Stanford for this Datathon about Oil Palm Plantations !
https://www.widsconference.org/datathon-details.html
https://www.kaggle.com/c/widsdatathon2019

See below our summary solution:
https://www.kaggle.com/c/widsdatathon2019/discussion/82284
We would like to first thanks WiDS and all the sponsors and Kaggle for providing us with an interesting use case that might have a good inpact for our planet .

We are currently coworkers and just after participating to the Kaggle days in Paris we decided that we should team up one day, this challenge was obviously a good opportunity (never mind if we miss the xbox at the end).

Solution:

  • we started with the public kernel https://www.kaggle.com/tcapelle/fastai-starter based on fast.ai , thanks @tcapelle for that one.
    • we then follow this check list of things to try :
      • Stratify the cross validation on the target, eventually also on the score column
      • Lower the % of validation, we ended up with a 7 and 11 fold cross validation
      • Try different image size , we ended up using 256x256 image size
      • Try different kind of pretrained model, SeResnext101 and Senet154 were those that gave the best results. You can used this repo to enrich your fast.ai experience with lot of models : https://github.com/Cadene/pretrained-models.pytorch
      • Tune the number of epochs and LR
      • Use the default data augmentation and TTA provided by fast.ai
      • We finally bagged using the median of the results from all the fold of each models
      • We did most of the processing thru kaggle kernel or google colab notebook.

What we have tried but didn’t improve :

  • data augmentation with SMOTE
  • other kind of transformations for the fast.ai transform function
  • overtuning the network
  • psuedo labelling

the final kernel is still running we will publish it when it’s done.

reference:


https://www.fast.ai/

voila/

Lu, Marie-Angèle, Jérôme et Mathurin

4 Likes

Congratulations on the win! :slight_smile:

How to use SEResNext101 with fastai ? Is there any example code ?

You can find an example here
https://www.kaggle.com/jayjay75/first-place-solution

2 Likes

It’s super. Thanks a lot :slight_smile: